Readonly / Non-deduplicating Mode

When run without -d (the default) duperemove will print out one or more tables of matching extents it has determined would be ideal candidates for deduplication. As a result, readonly mode is useful for seeing what duperemove might do when run with -d.

Generally, duperemove does not concern itself with the underlying representation of the extents it processes. Some of them could be compressed, undergoing I/O, or even have already been deduplicated. In dedupe mode, the kernel handles those details and therefore we try not to replicate that work.

Deduping Mode

This functions similarly to readonly mode with the exception that the duplicated extents found in our “read, hash, and compare” step will actually be submitted for deduplication. Extents that have already been deduped will be skipped. An estimate of the total data deduplicated will be printed after the operation is complete. This estimate is calculated by comparing the total amount of shared bytes in each file before and after the dedupe.

Common options

files can refer to a list of regular files and directories or be a hyphen (-) to read them from standard input. If a directory is specified, all regular files within it will also be scanned. Duperemove can also be told to recursively scan directories with the -r switch.

-r

Enable recursive dir traversal.

-d

De-dupe the results - only works on btrfs and xfs. Use this option twice to disable the check and try to run the ioctl anyway.

--hashfile=hashfile

Use a file for storage of hashes instead of memory. This option drastically reduces the memory footprint of duperemove and is recommended when your data set is more than a few files large. Hashfiles are also reusable, allowing you to further reduce the amount of hashing done on subsequent dedupe runs.

-B N, --batchsize=N

Run the deduplication phase every N files newly scanned. This greatly reduces memory usage for large dataset, or when you are doing partial extents lookup, but reduces multithreading efficiency.

-h

Print numbers in human-readable format.

-q

Quiet mode. Duperemove will only print errors and a short summary of any dedupe.

-v

Be verbose.

--help

Prints help text.

Advanced options

--fdupes

Run in fdupes mode. With this option you can pipe the output of fdupes to duperemove to dedupe any duplicate files found. When receiving a file list in this manner, duperemove will skip the hashing phase.

-L

Print all files in the hashfile and exit. Requires the --hashfile option. Will print additional information about each file when run with -v.

-R files ..

Remove file from the db and exit. Duperemove will read the list from standard input if a hyphen (-) is provided. Requires the --hashfile option.

--skip-zeroes

Read data blocks and skip any zeroed blocks, useful for speedup duperemove, but can prevent deduplication of zeroed files.

-b size

Use the specified block size for reading file extents. Defaults to 128K.

--io-threads=N

Use N threads for I/O. This is used by the file hashing and dedupe stages. Default is automatically detected based on number of host cpus.

--cpu-threads=N

Use N threads for CPU bound tasks. This is used by the duplicate extent finding stage. Default is automatically detected based on number of host cpus.

--dedupe-options=options

Comma separated list of options which alter how we dedupe. Prepend `no' to an option in order to turn it off.

--read-hashes=hashfile

This option is primarily for testing. See the --hashfile option if you want to use hashfiles.

--write-hashes=hashfile

This option is primarily for testing. See the --hashfile option if you want to use hashfiles.

--debug

Print debug messages, forces -v if selected.

--hash-threads=N

Deprecated, see --io-threads above.

--exclude=PATTERN

You can exclude certain files and folders from the deduplication process. This might be benefical for skipping subvolume snapshot mounts, for instance. Unless you provide a full path for exclusion, the exclude will be relative to the current working directory. Another thing to keep in mind is that shells usually expand glob pattern so the passed in pattern ought to also be quoted. Taking everything into consideration the correct way to pass an exclusion pattern is duperemove --exclude "/path/to/dir/file*" /path/to/dir

Simple Usage

Dedupe the files in directory /foo, recurse into all subdirectories. You only want to use this for small data sets:

duperemove -dr /foo

Use duperemove with fdupes to dedupe identical files below directory foo:

fdupes -r /foo | duperemove --fdupes

Using Hashfiles

Duperemove can optionally store the hashes it calculates in a hashfile. Hashfiles have two primary advantages - memory usage and re-usability. When using a hashfile, duperemove will stream computed hashes to it, instead of main memory.

If Duperemove is run with an existing hashfile, it will only scan those files which have changed since the last time the hashfile was updated. The files argument controls which directories duperemove will scan for newly added files. In the simplest usage, you rerun duperemove with the same parameters and it will only scan changed or newly added files - see the first example below.

Dedupe the files in directory foo, storing hashes in foo.hash. We can run this command multiple times and duperemove will only checksum and dedupe changed or newly added files:

duperemove -dr --hashfile=foo.hash foo/

Don’t scan for new files, only update changed or deleted files, then dedupe:

duperemove -dr --hashfile=foo.hash

Add directory bar to our hashfile and discover any files that were recently added to foo:

duperemove -dr --hashfile=foo.hash foo/ bar/

List the files tracked by foo.hash:

duperemove -L --hashfile=foo.hash

Is duperemove safe for my data?

Yes. To be specific, duperemove does not deduplicate the data itself. It simply finds candidates for dedupe and submits them to the Linux kernel FIDEDUPERANGE ioctl. In order to ensure data integrity, the kernel locks out other access to the file and does a byte-by-byte compare before proceeding with the dedupe.

Is is safe to interrupt the program (Ctrl-C)?

Yes. The Linux kernel deals with the actual data. On Duperemove’ side, a transactional database engine is used. The result is that you should be able to ctrl-c the program at any point and re-run without experiencing corruption of your hashfile. In case of a bug, your hashfile may be broken, but your data never will.

I got two identical files, why are they not deduped?

Duperemove by default works on extent granularity. What this means is if there are two files which are logically identical (have the same content) but are laid out on disk with different extent structure they won’t be deduped. For example if 2 files are 128k each and their content are identical but one of them consists of a single 128k extent and the other of 2 * 64k extents then they won’t be deduped. This behavior is dependent on the current implementation and is subject to change as duperemove is being improved.

What is the cost of deduplication?

Deduplication will lead to increased fragmentation. The blocksize chosen can have an effect on this. Larger blocksizes will fragment less but may not save you as much space. Conversely, smaller block sizes may save more space at the cost of increased fragmentation.

How can I find out my space savings after a dedupe?

Duperemove will print out an estimate of the saved space after a dedupe operation for you.

You can get a more accurate picture by running `btrfs fi df' before and after each duperemove run.

Be careful about using the `df' tool on btrfs - it is common for space reporting to be `behind' while delayed updates get processed, so an immediate df after deduping might not show any savings.

Why is the total deduped data report an estimate?

At the moment duperemove can detect that some underlying extents are shared with other files, but it can not resolve which files those extents are shared with.

Imagine duperemove is examining a series of files and it notes a shared data region in one of them. That data could be shared with a file outside of the series. Since duperemove can’t resolve that information it will account the shared data against our dedupe operation while in reality, the kernel might deduplicate it further for us.

Why are my files showing dedupe but my disk space is not shrinking?

This is a little complicated, but it comes down to a feature in Btrfs called bookending. The Btrfs wiki (http://en.wikipedia.org/wiki/Btrfs#Extents) explains this in detail.

Essentially though, the underlying representation of an extent in Btrfs can not be split (with small exception). So sometimes we can end up in a situation where a file extent gets partially deduped (and the extents marked as shared) but the underlying extent item is not freed or truncated.

Is there an upper limit to the amount of data duperemove can process?

Duperemove is fast at reading and cataloging data. Dedupe runs will be memory limited unless the --hashfile option is used. --hashfile allows duperemove to temporarily store duplicated hashes to disk, thus removing the large memory overhead and allowing for a far larger amount of data to be scanned and deduped. Realistically though you will be limited by the speed of your disks and cpu. In those situations where resources are limited you may have success by breaking up the input data set into smaller pieces.

When using a hashfile, duperemove will only store duplicate hashes in memory. During normal operation then the hash tree will make up the largest portion of duperemove memory usage. As of Duperemove v0.11 hash entries are 88 bytes in size. If you know the number of duplicate blocks in your data set you can get a rough approximation of memory usage by multiplying with the hash entry size.

Actual performance numbers are dependent on hardware - up to date testing information is kept on the duperemove wiki (see below for the link).

How large of a hashfile will duperemove create?

Hashfiles are essentially sqlite3 database files with several tables, the largest of which are the files and extents tables. Each extents table entry is about 72 bytes though that may grow as features are added. The size of a files table entry depends on the file path but a good estimate is around 270 bytes per file. The number of extents in a data set is directly proportional to file fragmentation level.

If you know the total number of extents and files in your data set then you can calculate the hashfile size as:

Hashfile Size = Num Hashes * 72 + Num Files * 270

Using a real world example of 1TB (8388608 128K blocks) of data over 1000 files:

8388608 * 72 + 270 * 1000 = 755244720 or about 720MB for 1TB spread over 1000 files.

Note that none of this takes database overhead into account.